Digital intelligent live broadcast bullet screen supervision method and system

Abstract

The invention relates to the technical field of live broadcast management, in particular to a digital intelligent live broadcast bullet screen supervision method and system. The method comprises the following steps: acquiring live broadcast bullet screen data and real-time live broadcast data; analyzing the live broadcast bullet screen data, and determining an abnormal bullet screen and an abnormal level of the abnormal bullet screen; performing risk assessment on the abnormal bullet screen in combination with the real-time live broadcast data, and determining a dynamic shielding judgment threshold value of the abnormal bullet screen; and comparing the abnormal level with the dynamic shielding judgment threshold, and determining a display strategy of the abnormal bullet screen according to a level comparison result. And in combination with the real-time live broadcast data, risk assessment is performed on the abnormal bullet screen, the dynamic shielding judgment threshold of the abnormal bullet screen is determined, and the high-risk scene recognition accuracy is improved. And comparing the abnormal level with a dynamic shielding judgment threshold, and determining a display strategy of the abnormal bullet screen according to a level comparison result, thereby reducing interference, retaining free expression of the user, reducing the complaint rate of the user, and inhibiting malicious screen refreshing diffusion.

Description

Technical Field

[0001] The present application relates to the field of live broadcast management technology, and in particular to a digital intelligent live broadcast barrage supervision method and system. Background Art

[0002] With the rapid development of live streaming, real-time interactive commentary has become an important way for users to engage with live content. However, the immediacy and anonymity of commentary also lead to issues such as malicious comments, cyberbullying, and excessive memes.

[0003] Offensive barrage comments and frequent screen-sweeping malicious memes not only disrupt the viewing experience but can also trigger mass negative public opinion. Existing live streaming platforms generally use blocking mechanisms based on static sensitive word libraries, which make it difficult to distinguish contextual differences and promptly intercept malicious screen-sweeping behavior. Summary of the Invention

[0004] This application provides a digital intelligent live broadcast barrage supervision method and system to solve the above problems.

[0005] In a first aspect, the present application provides a digital intelligent live broadcast barrage supervision method, the method comprising: Acquire live broadcast barrage data and real-time live broadcast data; analyze the live broadcast barrage data to determine abnormal barrages and the abnormality level of the abnormal barrages; In combination with the real-time live broadcast data, a risk assessment is performed on the abnormal barrage, and a dynamic blocking determination threshold for the abnormal barrage is determined; The abnormality level is compared with the dynamic shielding determination threshold, and the display strategy of the abnormal barrage is determined according to the level comparison result.

[0006] This solution captures live barrage data and real-time live broadcast data, avoiding semantic fragmentation caused by single-modal processing. Analyze live barrage data to identify abnormal barrages and their abnormality levels, achieving objective quantification of risk levels and avoiding subjective biases in manual rules. Combined with real-time live broadcast data, conduct a risk assessment of abnormal barrages, determine the dynamic blocking threshold for abnormal barrages, and improve the accuracy of identifying high-risk scenarios. Compare the abnormality level with the dynamic blocking threshold. Based on the level comparison results, determine the display strategy for abnormal barrages, reducing interference while preserving user freedom of expression, lowering user complaint rates, and suppressing the spread of malicious screen swiping.

[0007] Optionally, analyzing the live broadcast barrage data to determine abnormal barrages and the abnormality level of the abnormal barrages includes: Analyze the real-time live broadcast data to determine the live broadcast type; Based on the live broadcast type, obtain a preset sensitive word library; Based on the preset sensitive word library, analyzing the live broadcast barrage data to determine abnormal barrage; Analyze the abnormal comments and determine the degree of deviation of values; The abnormality level of the abnormal barrage is determined according to the value deviation.

[0008] Through this solution, real-time live broadcast data is analyzed to determine the type of live broadcast, eliminating the problem of misjudgment caused by the lack of scene features. Based on the live broadcast type, a preset sensitive word library is obtained. While retaining the basic sensitive word library, the false blocking rate of normal interactive barrage is reduced, thereby improving scene adaptability. Based on the preset sensitive word library, live broadcast barrage data is analyzed to determine abnormal barrages and improve the recognition accuracy of offensive barrages. Abnormal barrages are analyzed to determine the value deviation degree, quantify the degree of harm of the barrage values, and avoid the binary shielding mechanism from improperly handling gray area content. According to the value deviation degree, the abnormal level of abnormal barrage is determined to implement risk classification management and control, while reducing the false injury rate and improving the response speed to high-risk barrages.

[0009] Optionally, analyzing the live broadcast barrage data based on the preset sensitive word library to determine abnormal barrage includes: Based on the real-time live broadcast data, historical live broadcast barrage data is obtained; Determining malicious semantic deformation features based on the preset sensitive word library and the historical live broadcast barrage data; Constructing a malicious semantic feature matrix based on the malicious semantic deformation features; Mapping the live broadcast barrage data to the malicious semantic feature matrix through the BERT model to obtain a multi-dimensional matrix value; Abnormal comments are determined according to the multi-dimensional matrix value.

[0010] Through this solution, based on real-time live broadcast data, historical live broadcast barrage data is obtained to eliminate redundant noise, ensure the adaptability of feature extraction to the current live broadcast content, and avoid cross-scene interference. According to the preset sensitive word library and historical live broadcast barrage data, malicious semantic deformation features are determined to improve the coverage of hidden malicious text, exclude high-frequency but harmless deformation words, reduce the false positive rate, and retain features that are strongly associated with malicious semantics. Based on the malicious semantic deformation features, a malicious semantic feature matrix is ​​constructed to eliminate weight skew caused by magnitude differences and ensure the fairness of the matrix mapping results. The live broadcast barrage data is mapped to the malicious semantic feature matrix through the BERT model to obtain multi-dimensional matrix values, which helps to enhance the semantic analysis ability of obscure variant words and convert text risks into quantifiable indicators. According to the multi-dimensional matrix values, abnormal barrage is determined, which helps to eliminate the threshold rigidity problem and alleviate the feedback lag defect.

[0011] Optionally, analyzing the abnormal comments and determining the value deviation includes: Obtaining historical cyber violence events, analyzing the historical cyber violence events, and determining text feature distribution patterns; According to the live broadcast type, obtain the supervision level setting parameters of the current live broadcast room; Determine a preset multi-dimensional warning threshold based on the text feature distribution law and the supervision level setting parameters; The multi-dimensional matrix value of the abnormal barrage is compared with the preset multi-dimensional warning threshold, and the value deviation is determined according to the warning comparison result.

[0012] This solution captures and analyzes historical online violence incidents, identifies the distribution patterns of text features, and overcomes the flaw of static lexicons in capturing dynamic semantic combination risks, providing a fine-grained, scenario-based basis for threshold setting. Based on the live broadcast type, the regulatory level setting parameters for the current live broadcast room are obtained to suppress the missed detection rate of semantic deformation attacks in highly sensitive scenarios and avoid accidental injuries or missed detections caused by a one-size-fits-all threshold. Based on the distribution patterns of text features and regulatory level setting parameters, a preset multi-dimensional warning threshold is determined to eliminate the problem of delayed response during peak screen-swiping periods and adapt to the risk fluctuations during the live broadcast process. The multi-dimensional matrix values ​​of abnormal comments are compared with the preset multi-dimensional warning thresholds. Based on the warning comparison results, the value deviation is determined. This helps overcome the blind spot of binary judgment in identifying complex malicious expressions and improves the accuracy of handling mixed comments such as "memes + insults".

[0013] Optionally, the step of performing a risk assessment on the abnormal barrage in combination with the real-time live broadcast data and determining a threshold for dynamic shielding of the abnormal barrage includes: Analyze the abnormal comments according to the malicious semantic feature matrix to determine variant text; Analyze the real-time live broadcast data to determine the live broadcast image characteristics at the current moment; Acquire historical real-time live broadcast data of the current live broadcast room from the start of the live broadcast to the current moment, analyze the historical real-time live broadcast data, and determine the distribution characteristics of the abnormal barrage and the variant text barrage; Determining the bullet screen trigger weight of the live broadcast screen feature according to the bullet screen distribution feature; The barrage trigger weight is used to perform a risk assessment on the abnormal barrage, and a dynamic shielding judgment threshold for the abnormal barrage is determined.

[0014] This solution analyzes abnormal barrages and identifies variant text based on a malicious semantic feature matrix, helping to eliminate semantic escapes caused by the inability of the sensitive word library to cover phono-semantic character substitutions and improving the accuracy of identifying variant sensitive words. Real-time live broadcast data is analyzed to determine the current live broadcast screen features, avoiding the poor scene adaptability caused by single text features. Historical real-time live broadcast data from the start of the current live broadcast to the current moment is obtained and analyzed to determine the barrage distribution features of abnormal barrages and variant text, helping to eliminate the defect of missing multimodal associations. Based on the barrage distribution features, the barrage trigger weights for live broadcast screen features are determined, breaking through the limitations of threshold rigidity and feedback lag. Using the barrage trigger weights, risk assessments are performed on abnormal barrages, and a dynamic blocking threshold for abnormal barrages is determined to avoid an increase in the misjudgment rate due to user fatigue during long-term live broadcasts.

[0015] Optionally, determining the bullet screen triggering weight of the live broadcast screen feature according to the bullet screen distribution feature includes: Determining the frequency of occurrence of the abnormal barrage and the variant text in the corresponding distribution screen according to the barrage distribution characteristics; Generate scene attention weights based on the occurrence frequencies; Performing a convolution operation on the distribution feature and the scene attention weight to output a dynamic baseline value; Based on the dynamic baseline value, the live broadcast screen features are analyzed to determine the bullet screen triggering weight of the live broadcast screen features.

[0016] This solution determines the frequency of occurrence of abnormal barrages and variant text in the corresponding distributed screen based on the barrage distribution characteristics, helping to quantify the spatiotemporal distribution density of different areas of the screen. Based on the frequency of occurrence, a scene attention weight is generated to help characterize the risk level of malicious content in the current scene. Convolution operations are performed on the distribution characteristics and the scene attention weight to output a dynamic baseline value, which helps to respond to feedback lag optimization needs. Based on the dynamic baseline value, the live screen characteristics are analyzed and the barrage trigger weights of the live screen characteristics are determined, which helps to achieve the core goal of adaptive risk management.

[0017] Optionally, the utilizing the barrage trigger weight to perform risk assessment on the abnormal barrage and determine a dynamic shielding determination threshold for the abnormal barrage includes: Determining an initial tolerance threshold according to the live broadcast type; Determining the rationality of the occurrence of the abnormal barrage based on the barrage trigger weight; Obtaining historical reporting data, analyzing the historical reporting data, and determining a negative feedback threshold; According to the rationality of the occurrence and the negative feedback threshold, the initial tolerance threshold is adjusted to obtain the dynamic shielding determination threshold of the abnormal barrage.

[0018] This solution determines the initial tolerance threshold based on the live broadcast type, helping to eliminate the problem of poor scene adaptability and avoid misjudgments caused by a one-size-fits-all approach. Determining the legitimacy of abnormal barrages based on their trigger weights helps quickly suppress malicious screen-swiping behavior and avoid excessive interference with benign interactions. Historical reporting data is obtained and analyzed to determine the negative feedback threshold and block the risk of negative content spreading. Adjusting the initial tolerance threshold based on the legitimacy of occurrence and the negative feedback threshold yields a dynamic blocking threshold for abnormal barrages, helping to maintain basic functionality in extreme situations and avoid the risk of system crashes caused by delayed feedback.

[0019] Optionally, analyzing the live screen feature based on the dynamic baseline value to determine the bullet screen triggering weight of the live screen feature includes: Analyze the live broadcast image features to determine visual feature vectors and audio keyword vectors; Performing cross-modal attention fusion on the visual feature vector and the audio keyword vector to generate a scene semantic vector; Parsing the abnormal comment and the variant text to determine a text semantic vector; Calculating the dynamic cosine similarity between the text semantic vector and the scene semantic vector; The dynamic cosine similarity is matched with the dynamic baseline value, and the bullet screen triggering weight of the live screen feature is determined according to the matching result.

[0020] This solution analyzes live video features to determine visual feature vectors and audio keyword vectors, helping to eliminate the lack of correlation between video features and comment content. It also captures the potential correlation between audio semantics and comment content, avoiding misjudgments caused by independently processing modal data. Cross-modal attention fusion of visual feature vectors and audio keyword vectors generates scene semantic vectors, helping to establish a joint semantic space for video, audio, and comment content, providing a baseline scene context for dynamic similarity calculation. It analyzes anomalous comment content and variant text to determine text semantic vectors, capturing implicit malicious intent in comment content and avoiding mechanical misjudgments caused by keyword matching. It calculates the dynamic cosine similarity between text semantic vectors and scene semantic vectors, dynamically identifying scene mismatch risks and eliminating poor scene adaptability. The dynamic cosine similarity is matched with a dynamic baseline value. Based on the matching results, the comment trigger weight for live video features is determined, enabling dynamic control of threshold rigidity and feedback lag, rapidly improving blocking sensitivity.

[0021] Optionally, constructing a malicious semantic feature matrix according to the malicious semantic deformation feature includes: Simulating the semantic derivation path of the malicious semantic deformation feature through a generative adversarial network; According to the semantic derivation path, a malicious semantic feature matrix is ​​constructed.

[0022] This solution uses a generative adversarial network to simulate the semantic derivation path of malicious semantic deformation features, providing a mathematical representation of the dynamic deformation trajectory for feature matrix construction and supporting the abstract extraction of semantic deformation patterns. Constructing a malicious semantic feature matrix based on the semantic derivation path helps meet the technical requirements for semantic deformation feature mining.

[0023] In a second aspect, the present application provides a digital intelligent live broadcast barrage monitoring system, the system comprising: A data analysis module is used to obtain live broadcast barrage data and real-time live broadcast data; analyze the live broadcast barrage data, determine abnormal barrages and the abnormality level of the abnormal barrages; A threshold determination module is used to perform a risk assessment on the abnormal barrage in combination with the real-time live broadcast data, and determine a dynamic shielding determination threshold for the abnormal barrage; The strategy determination module is used to compare the abnormal level with the dynamic shielding judgment threshold, and determine the display strategy of the abnormal barrage according to the level comparison result.

[0024] Optionally, when the data analysis module analyzes the live broadcast barrage data and determines abnormal barrages and the abnormality level of the abnormal barrages, it is used to: Analyze the real-time live broadcast data to determine the live broadcast type; Based on the live broadcast type, obtain a preset sensitive word library; Based on the preset sensitive word library, analyzing the live broadcast barrage data to determine abnormal barrage; Analyze the abnormal comments and determine the degree of deviation of values; The abnormality level of the abnormal barrage is determined according to the value deviation.

[0025] Optionally, the data analysis module analyzes the live broadcast barrage data based on the preset sensitive word library, and when determining abnormal barrage, is used to: Based on the real-time live broadcast data, historical live broadcast barrage data is obtained; Determining malicious semantic deformation features based on the preset sensitive word library and the historical live broadcast barrage data; Constructing a malicious semantic feature matrix based on the malicious semantic deformation features; Mapping the live broadcast barrage data to the malicious semantic feature matrix through the BERT model to obtain a multi-dimensional matrix value; Abnormal comments are determined according to the multi-dimensional matrix value.

[0026] Optionally, when the data analysis module analyzes the abnormal comments and determines the value deviation, it is used to: Obtaining historical cyber violence events, analyzing the historical cyber violence events, and determining text feature distribution patterns; According to the live broadcast type, obtain the supervision level setting parameters of the current live broadcast room; Determine a preset multi-dimensional warning threshold based on the text feature distribution law and the supervision level setting parameters; The multi-dimensional matrix value of the abnormal barrage is compared with the preset multi-dimensional warning threshold, and the value deviation is determined according to the warning comparison result.

[0027] Optionally, the threshold determination module combines the real-time live broadcast data to perform a risk assessment on the abnormal barrage, and when determining a dynamic shielding determination threshold for the abnormal barrage, is used to: Analyze the abnormal comments according to the malicious semantic feature matrix to determine variant text; Analyze the real-time live broadcast data to determine the live broadcast image characteristics at the current moment; Acquire historical real-time live broadcast data of the current live broadcast room from the start of the live broadcast to the current moment, analyze the historical real-time live broadcast data, and determine the distribution characteristics of the abnormal barrage and the variant text barrage; Determining the bullet screen trigger weight of the live broadcast screen feature according to the bullet screen distribution feature; The barrage trigger weight is used to perform a risk assessment on the abnormal barrage, and a dynamic shielding judgment threshold for the abnormal barrage is determined.

[0028] Optionally, when the threshold determination module determines the bullet screen trigger weight of the live screen feature according to the bullet screen distribution feature, it is used to: Determining the frequency of occurrence of the abnormal barrage and the variant text in the corresponding distribution screen according to the barrage distribution characteristics; Generate scene attention weights based on the occurrence frequencies; Performing a convolution operation on the distribution feature and the scene attention weight to output a dynamic baseline value; Based on the dynamic baseline value, the live broadcast screen features are analyzed to determine the bullet screen triggering weight of the live broadcast screen features.

[0029] Optionally, the threshold determination module uses the barrage triggering weight to perform a risk assessment on the abnormal barrage, and when determining the dynamic shielding determination threshold for the abnormal barrage, is used to: Determining an initial tolerance threshold according to the live broadcast type; Determining the rationality of the occurrence of the abnormal barrage based on the barrage trigger weight; Obtaining historical reporting data, analyzing the historical reporting data, and determining a negative feedback threshold; According to the rationality of the occurrence and the negative feedback threshold, the initial tolerance threshold is adjusted to obtain the dynamic shielding determination threshold of the abnormal barrage.

[0030] Optionally, when the threshold determination module analyzes the live screen features based on the dynamic baseline value and determines the bullet screen triggering weight of the live screen features, it is configured to: Analyze the live broadcast image features to determine visual feature vectors and audio keyword vectors; Performing cross-modal attention fusion on the visual feature vector and the audio keyword vector to generate a scene semantic vector; Parsing the abnormal comment and the variant text to determine a text semantic vector; Calculating the dynamic cosine similarity between the text semantic vector and the scene semantic vector; The dynamic cosine similarity is matched with the dynamic baseline value, and the bullet screen triggering weight of the live screen feature is determined according to the matching result.

[0031] Optionally, when constructing a malicious semantic feature matrix based on the malicious semantic deformation feature, the data analysis module is used to: Simulating the semantic derivation path of the malicious semantic deformation feature through a generative adversarial network; According to the semantic derivation path, a malicious semantic feature matrix is ​​constructed. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0033] Figure 1 A schematic diagram of an application scenario provided for an embodiment of the present application.

[0034] Figure 2 A flowchart of a digital intelligent live broadcast barrage supervision method provided in one embodiment of the present application.

[0035] Figure 3 A structural diagram of a digital intelligent live broadcast barrage supervision system provided in one embodiment of the present application. DETAILED DESCRIPTION

[0036] To make the purpose, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0037] In this document, the term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document, unless otherwise specified, generally indicates an "or" relationship between the related objects.

[0038] The embodiments of the present application are described in further detail below with reference to the accompanying drawings.

[0039] Offensive barrage comments and frequent screen-sweeping malicious memes not only disrupt the viewing experience but can also trigger mass negative public opinion. Existing live streaming platforms generally use blocking mechanisms based on static sensitive word libraries, which struggle to distinguish between contextual differences. Relying on passive response models like manual review or user reporting makes it difficult to promptly intercept malicious screen-sweeping behavior.

[0040] Based on this, the present application provides a digital intelligent live broadcast barrage supervision method and system, which obtains live broadcast barrage data and real-time live broadcast data; analyzes the live broadcast barrage data to determine abnormal barrages and the abnormal level of the abnormal barrages; combines the real-time live broadcast data to conduct a risk assessment on the abnormal barrages and determine the dynamic shielding judgment threshold of the abnormal barrages; compares the abnormal level with the dynamic shielding judgment threshold, and determines the display strategy of the abnormal barrage based on the level comparison result. Live broadcast barrage data and real-time live broadcast data are obtained to avoid the semantic fragmentation problem caused by single-modal processing. Live broadcast barrage data is analyzed to determine abnormal barrages and the abnormal level of abnormal barrages, to achieve objective quantification of risk levels, and to avoid subjective bias of manual rules. Combined with real-time live broadcast data, a risk assessment is conducted on abnormal barrages, and a dynamic shielding judgment threshold for abnormal barrages is determined to improve the accuracy of high-risk scene recognition. The abnormal level is compared with the dynamic shielding judgment threshold, and based on the level comparison result, a display strategy for abnormal barrages is determined to reduce interference while retaining user freedom of expression, reduce user complaint rates, and suppress the spread of malicious screen swiping.

[0041] Figure 1A schematic diagram of an application scenario provided by this application, when performing live broadcast management, the method provided by this application is applied. Specifically, the method provided by this application is applied to any server, the server interacts with the live broadcast device, and the live broadcast barrage data and real-time live broadcast data are obtained and analyzed through the live broadcast device to determine abnormal barrages and abnormal levels of abnormal barrages. Combined with real-time live broadcast data, a risk assessment is conducted on abnormal barrages, and the dynamic shielding judgment threshold of abnormal barrages is determined to improve the accuracy of high-risk scene recognition. The abnormal level is compared with the dynamic shielding judgment threshold, and according to the level comparison results, the display strategy of abnormal barrages is determined to reduce interference while retaining user freedom of expression, reduce user complaint rates, and suppress the spread of malicious screen swiping.

[0042] For specific implementation methods, please refer to the following embodiments.

[0043] Figure 2 This is a flowchart of a digital intelligent live broadcast bullet screen monitoring method provided in an embodiment of the present application. The method of this embodiment can be applied to the server in the above scenario. Figure 2 As shown, the method includes: S201. Obtain live broadcast barrage data and real-time live broadcast data; analyze the live broadcast barrage data to determine abnormal barrages and abnormal levels of abnormal barrages; Live broadcast barrage data can be the barrage text stream data sent by users in real time in the live broadcast room.

[0044] Real-time live broadcast data can be a multimodal data set directly associated with the live broadcast content.

[0045] Abnormal barrages may be barrage texts that contain malicious comments, cyberbullying, or excessive meme-related content.

[0046] The abnormality level may be a barrage risk level.

[0047] Specifically, live broadcast barrage data is received in real time through the live broadcast platform API; real-time live broadcast data aligned with the barrage timeline is synchronously collected. Then, based on the generative adversarial network framework, an adversarial sample generation model is constructed by iteratively generating adversarial text variants that are semantically equivalent to the original sensitive words and verifying the effectiveness of the perturbation. Furthermore, a dynamic variant sensitive word library is constructed through the adversarial sample generation model to identify semantic deformation features such as homophone replacement, pinyin abbreviations, and emoticon insertion. A scenario-based semantic whitelist is established by combining the live broadcast room classification labels with real-time interactive events. When the live broadcast barrage data contains product-related words, it is detected whether there is a product display area in the live broadcast screen. If there is no product in the screen and the live broadcast barrage data contains variant insults, an abnormal barrage is determined. Then, based on the feature importance analysis method of historical abnormal sample statistics, the positive correlation strength of different deformation features and abnormal barrage in the historical data is statistically analyzed to quantify the weight of the semantic deformation feature and generate an abnormal confidence score, thereby determining the abnormal level of the abnormal barrage.

[0048] S202. Conduct a risk assessment on abnormal barrages based on real-time live broadcast data, and determine a dynamic blocking threshold for abnormal barrages. The dynamic shielding determination threshold may be a dynamically adjusted risk determination boundary value.

[0049] Specifically, based on the transfer learning theory, the target detection model is used for domain adaptive fine-tuning, and the bounding box regression and anchor mechanism are used to perform multi-scale object detection on the main object in the live broadcast screen; then, the host's voice and intonation features are extracted to judge the current emotional state; then, the screen text is recognized by OCR to establish a time anchor; then, based on the vector space model theory, the barrage text and scene features are mapped to the same semantic space through word embedding, and the cross-modal semantic association strength is measured using cosine similarity to construct a barrage-scene association matrix; the semantic matching degree between the barrage keywords and the current picture / voice is judged; then, the standard deviation of the frequency of similar barrages in a unit time window and the historical baseline is counted; then, the degree of deviation between the text emotional tendency and the host's voice emotion is detected; finally, the dynamic shielding judgment threshold of abnormal barrages is determined based on the correlation calculation results.

[0050] S203: Compare the abnormality level with the dynamic shielding judgment threshold, and determine the display strategy of the abnormal bullet comments based on the level comparison result.

[0051] The level comparison result may be the difference between the abnormal level of the abnormal bullet comment and the dynamic shielding determination threshold.

[0052] The display strategy can be a barrage control measure.

[0053] Specifically, the abnormality level is compared with the dynamic shielding judgment threshold. If the abnormality level is higher than the dynamic shielding judgment threshold, the barrage is completely filtered; if the abnormality level is between the dynamic shielding judgment threshold, it is added to the delay queue and released at a controllable rate; if the abnormality level is lower than the dynamic shielding judgment threshold but there is a potential risk, a visual mark is added to the barrage side; finally, based on the comparison results of different levels, the display strategy of abnormal barrage is determined.

[0054] This solution captures live barrage data and real-time live broadcast data, avoiding semantic fragmentation caused by single-modal processing. Analyze live barrage data to identify abnormal barrages and their abnormality levels, achieving objective quantification of risk levels and avoiding subjective biases in manual rules. Combined with real-time live broadcast data, conduct a risk assessment of abnormal barrages, determine the dynamic blocking threshold for abnormal barrages, and improve the accuracy of identifying high-risk scenarios. Compare the abnormality level with the dynamic blocking threshold. Based on the level comparison results, determine the display strategy for abnormal barrages, reducing interference while preserving user freedom of expression, lowering user complaint rates, and suppressing the spread of malicious screen swiping.

[0055] In some embodiments, real-time live broadcast data is analyzed to determine the live broadcast type; based on the live broadcast type, a preset sensitive word library is obtained; based on the preset sensitive word library, live broadcast barrage data is analyzed to determine abnormal barrages; abnormal barrages are parsed to determine the value deviation; based on the value deviation, the abnormality level of the abnormal barrage is determined.

[0056] The live broadcast type may be a classification result that characterizes the core attributes of the current live broadcast content.

[0057] The preset sensitive word library can be a dynamic word library that is pre-set and bound to the live broadcast type, stored in the server in advance, and called when used.

[0058] Value deviation can be a comprehensive indicator to quantify the degree to which the content of the barrage deviates from positive values.

[0059] Specifically, a pre-set live broadcast classification model based on the ResNet-50 convolutional neural network architecture is called, leveraging its residual learning capabilities to perform multimodal feature fusion and classification training on live broadcast frame sequences and speech spectrogram features, thereby outputting the live broadcast type. Furthermore, based on the live broadcast type, a pre-set sensitive word library is called from the cloud-based policy library. Based on the pre-set sensitive word library, the barrage text is input into the adversarial sample generation model to generate a set of adversarial samples with homophone replacements, pinyin abbreviations, and emoticon insertions, thereby constructing a dynamic variant sensitive word library. Subsequently, combining the live broadcast type with real-time interactive events, the whitelist of words for the corresponding scenario is activated. Barrages containing product-related words are only detected when no product is displayed on the associated screen. When a barrage hits the dynamic variant sensitive word library, the target detection model is simultaneously called to verify whether there is a subject object in the live broadcast screen that matches the barrage semantics. If the image semantics conflict with the barrage, the barrage is marked as an anomaly. Next, for abnormal comments that hit dynamic variant sensitive words, an initial anomaly score is calculated based on the weighted semantic deformation features. The frequency deviation of similar comments within the time window is then added to generate a comprehensive anomaly confidence score. Furthermore, the abnormal comments are fed into a pre-trained value discrimination model based on the Transformer architecture. The natural language processing model is fine-tuned using a large-scale social ethics text dataset to extract features along three dimensions: semantic offensiveness, violation of social order and good morals, and malicious incitement. Finally, a basic level is assigned based on the anomaly confidence level. Higher values ​​deviations are assigned a higher level. Finally, the abnormal level of the abnormal comments is output based on the values ​​deviation level and the anomaly confidence level.

[0060] Through this solution, real-time live broadcast data is analyzed to determine the type of live broadcast, eliminating the problem of misjudgment caused by the lack of scene features. Based on the live broadcast type, a preset sensitive word library is obtained. While retaining the basic sensitive word library, the false blocking rate of normal interactive barrage is reduced, thereby improving scene adaptability. Based on the preset sensitive word library, live broadcast barrage data is analyzed to determine abnormal barrages and improve the recognition accuracy of offensive barrages. Abnormal barrages are analyzed to determine the value deviation degree, quantify the degree of harm of the barrage values, and avoid the binary shielding mechanism from improperly handling gray area content. According to the value deviation degree, the abnormal level of abnormal barrage is determined to implement risk classification management and control, while reducing the false injury rate and improving the response speed to high-risk barrages.

[0061] In some embodiments, based on real-time live broadcast data, historical live broadcast barrage data is obtained; malicious semantic deformation features are determined based on a preset sensitive word library and historical live broadcast barrage data; a malicious semantic feature matrix is ​​constructed based on the malicious semantic deformation features; the live broadcast barrage data is mapped to the malicious semantic feature matrix through a BERT model to obtain a multi-dimensional matrix value; and abnormal barrage is determined based on the multi-dimensional matrix value.

[0062] The historical live broadcast barrage data can be the barrage text and corresponding timestamp data generated by the target live broadcast room during the historical time period.

[0063] Malicious semantic deformation features may be a feature set used to identify semantic deformation attack behaviors.

[0064] The malicious semantic feature matrix can be a matrix data structure used to standardize the representation of semantic risk features.

[0065] The BERT model can be a pre-trained language model for mapping real-time live comment text into a high-dimensional semantic vector space.

[0066] The multi-dimensional matrix value may be a set of numerical values ​​used to quantify the abnormal risk level of the barrage.

[0067] Specifically, the target live stream's historical barrage text and corresponding timestamps are retrieved from the live streaming platform database, and the historical live barrage data is divided into preset time windows. Subsequently, based on the basic sensitive word list in the preset sensitive word library and the dynamic variant sensitive words output by the adversarial sample generation model, combined with the high-frequency semantic deformation patterns in the historical barrage data, malicious semantic deformation features are generated through word form-semantic association rule mining. Furthermore, the malicious semantic deformation features are divided into independent dimensions based on deformation type. Weights are statistically calculated based on the abnormal correlation strength of features in the historical barrage. A weight coefficient is then assigned to each dimension to form a malicious semantic feature matrix. The live barrage data is then fed into a BERT model based on a Transformer encoder architecture. Pre-trained using a masked language model and next sentence prediction tasks using large-scale unsupervised corpus, this BERT model is equipped to capture contextual semantic associations. It can directly load open-source pre-trained models to extract semantic vectors. Cosine similarity is used to calculate the matching value for each dimension of the malicious semantic feature matrix, outputting a multi-dimensional matrix value. Finally, a weighted threshold is set for the multi-dimensional matrix value. A barrage is classified as anomalous if at least two dimensions exceed the weighted threshold.

[0068] Through this solution, based on real-time live broadcast data, historical live broadcast barrage data is obtained to eliminate redundant noise, ensure the adaptability of feature extraction to the current live broadcast content, and avoid cross-scene interference. According to the preset sensitive word library and historical live broadcast barrage data, malicious semantic deformation features are determined to improve the coverage of hidden malicious text, exclude high-frequency but harmless deformation words, reduce the false positive rate, and retain features that are strongly associated with malicious semantics. Based on the malicious semantic deformation features, a malicious semantic feature matrix is ​​constructed to eliminate weight skew caused by magnitude differences and ensure the fairness of the matrix mapping results. The live broadcast barrage data is mapped to the malicious semantic feature matrix through the BERT model to obtain multi-dimensional matrix values, which helps to enhance the semantic analysis ability of obscure variant words and convert text risks into quantifiable indicators. According to the multi-dimensional matrix values, abnormal barrage is determined, which helps to eliminate the threshold rigidity problem and alleviate the feedback lag defect.

[0069] In some embodiments, historical cyber violence events are obtained and analyzed to determine the distribution pattern of text features; based on the live broadcast type, the regulatory level setting parameters of the current live broadcast room are obtained; based on the text feature distribution pattern and the regulatory level setting parameters, a preset multi-dimensional warning threshold is determined; the multi-dimensional matrix value of the abnormal barrage is compared with the preset multi-dimensional warning threshold, and the value deviation is determined based on the warning comparison result.

[0070] Historical online violence incidents can include malicious attacks or barrage text and associated data that violate community norms.

[0071] The text feature distribution law can be a quantitative distribution pattern of malicious semantic deformation features in terms of time, frequency, and co-occurrence relationship.

[0072] The current live broadcast room may be a virtual room instance that is currently running a live broadcast event.

[0073] The supervision level setting parameters may be rule parameters preset according to the live broadcast type for dynamically adjusting the risk determination criteria.

[0074] The preset multi-dimensional warning threshold can be a pre-set dynamic judgment boundary value for semantic deformation dimensions such as sound change, shape change, and symbol insertion. It is pre-stored in the server and called when used.

[0075] The warning comparison result can be a comparison status set of the multi-dimensional matrix value of the abnormal barrage and the warning threshold of the corresponding dimension.

[0076] Specifically, barrage samples associated with online violence incidents are extracted from historical live broadcast barrage data. A text feature distribution pattern is generated through word frequency statistics, co-occurrence word sequence mining, and time window density. Then, based on the classification label output by the live broadcast type, the regulatory level setting parameters for the current live broadcast room are loaded in real time. The statistical indicators in the text feature distribution pattern are weighted and fused with the regulatory level setting parameters to determine the preset multidimensional warning threshold: first, basic thresholds are set for sound change, shape change, and symbol insertion; then, the threshold offset is adjusted according to the regulatory level parameters; finally, the preset multidimensional warning threshold is determined. The multidimensional matrix values ​​of abnormal barrages are extracted and compared dimension by dimension with the preset multidimensional warning threshold. If the multidimensional matrix value exceeds the preset multidimensional warning threshold, it is marked as a trigger state. Then, based on the relative deviation measurement theory in probability statistics, the over-limit ratio is established by calculating the percentage by which the actual value of each trigger state exceeds the corresponding threshold. Subsequently, the number of trigger state dimensions and the over-limit ratio are counted to calculate the value deviation.

[0077] This solution captures and analyzes historical online violence incidents, identifies the distribution patterns of text features, and overcomes the flaw of static lexicons in capturing dynamic semantic combination risks, providing a fine-grained, scenario-based basis for threshold setting. Based on the live broadcast type, the regulatory level setting parameters for the current live broadcast room are obtained to suppress the missed detection rate of semantic deformation attacks in highly sensitive scenarios and avoid accidental injuries or missed detections caused by a one-size-fits-all threshold. Based on the distribution patterns of text features and regulatory level setting parameters, a preset multi-dimensional warning threshold is determined to eliminate the problem of delayed response during peak screen-swiping periods and adapt to the risk fluctuations during the live broadcast process. The multi-dimensional matrix values ​​of abnormal comments are compared with the preset multi-dimensional warning thresholds. Based on the warning comparison results, the value deviation is determined. This helps overcome the blind spot of binary judgment in identifying complex malicious expressions and improves the accuracy of handling mixed comments such as "memes + insults".

[0078] In some embodiments, based on the malicious semantic feature matrix, abnormal barrages are analyzed to determine variant texts; real-time live broadcast data are analyzed to determine the live broadcast screen features at the current moment; historical real-time live broadcast data of the current live broadcast room from the start of the live broadcast to the current moment is obtained, and the historical real-time live broadcast data is analyzed to determine the barrage distribution features of abnormal barrages and variant texts; based on the barrage distribution features, the barrage trigger weights of the live broadcast screen features are determined; the barrage trigger weights are used to perform risk assessments on abnormal barrages, and the dynamic shielding judgment threshold of abnormal barrages is determined.

[0079] The variant text may be a text obtained by semantically deforming a standard malicious word.

[0080] Live screen features can be dynamic visual attributes such as screen main objects, scene labels, text overlay area coordinates, and color contrast extracted from the live video stream.

[0081] Historical real-time live broadcast data can be a set of time series data accumulated from the start of the current live broadcast to the current moment.

[0082] The barrage distribution feature can be the quantitative distribution characteristics of abnormal barrage and variant text in the time and space dimensions.

[0083] The barrage trigger weight may be a parameter that characterizes the contribution of live screen features to the risk level of abnormal barrage.

[0084] Specifically, based on a malicious semantic feature matrix, the method calculates text vector similarity to match abnormal barrage comments with sound and shape change patterns in a pre-set deformation rule library. Then, it extracts text segments that meet the pre-set deformation rules to construct variant text. It analyzes live broadcast data and, using an image recognition model based on transfer learning theory and a pre-trained convolutional neural network as the feature extraction backbone, jointly optimizes the subject object and scene labels in the current image through a multi-task learning framework. It also simultaneously extracts the coordinates of the text overlay area and color contrast features, thereby generating the current live broadcast image features. It obtains historical live broadcast data from the start of the current broadcast to the current moment. Based on this historical live broadcast data, it counts the frequency of abnormal barrage comments at each stage of the live broadcast. It calculates the spatiotemporal alignment between the barrage triggering location and key image regions. It then establishes a mapping between the variant text type and the live broadcast scene to determine the barrage distribution characteristics of the variant text. Based on the barrage distribution characteristics and image features, it calculates the similarity between the current image label and the historical high-incidence scene labels. It then compares the deviation of the current variant text density from the historical distribution curve. Finally, it takes the weighted sum of the scene matching degree and the temporal deviation degree to generate the barrage trigger weight. The basic shielding threshold is set based on the supervision level parameters; the basic shielding threshold is dynamically corrected according to the maximum value in the barrage trigger weight matrix; finally, an attenuation factor is introduced according to the duration of the live broadcast to ultimately determine the dynamic shielding judgment threshold for abnormal barrage.

[0085] This solution analyzes abnormal barrages and identifies variant text based on a malicious semantic feature matrix, helping to eliminate semantic escapes caused by the inability of the sensitive word library to cover phono-semantic character substitutions and improving the accuracy of identifying variant sensitive words. Real-time live broadcast data is analyzed to determine the current live broadcast screen features, avoiding the poor scene adaptability caused by single text features. Historical real-time live broadcast data from the start of the current live broadcast to the current moment is obtained and analyzed to determine the barrage distribution features of abnormal barrages and variant text, helping to eliminate the defect of missing multimodal associations. Based on the barrage distribution features, the barrage trigger weights for live broadcast screen features are determined, breaking through the limitations of threshold rigidity and feedback lag. Using the barrage trigger weights, risk assessments are performed on abnormal barrages, and a dynamic blocking threshold for abnormal barrages is determined to avoid an increase in the misjudgment rate due to user fatigue during long-term live broadcasts.

[0086] In some embodiments, based on the barrage distribution characteristics, the frequency of occurrence of abnormal barrages and variant texts in the corresponding distribution screen is determined; based on the frequency of occurrence, the scene attention weight is generated; the distribution characteristics and the scene attention weight are convolved to output a dynamic baseline value; based on the dynamic baseline value, the live broadcast screen characteristics are analyzed to determine the barrage trigger weight of the live broadcast screen characteristics.

[0087] The corresponding distribution screens may be different areas divided in the live broadcast screen.

[0088] The occurrence frequency may be a statistical value of the number of times detected.

[0089] The scene attention weight can be a numerical value that represents the relative risk level of abnormal barrage distribution under a specific scene label.

[0090] The distribution feature can be a multidimensional matrix that describes the joint distribution state of abnormal comments in time, space and semantic dimensions.

[0091] The dynamic baseline value may be a dynamic reference value for a benchmark threshold value used to quantify the risk of barrage in the current scenario.

[0092] Specifically, based on the bullet-message distribution characteristics, the frequency of occurrence of abnormal bullet-messages and variant texts in the corresponding distribution screen within the preset time window is counted for the live broadcast screen. Then, the density of abnormal bullet-messages under each scene label is normalized according to the frequency of occurrence; then, the normalized value is mapped to the scene attention weight value through the Sigmoid activation function. The bullet-message distribution characteristics are constructed as a three-dimensional matrix of time window, screen area coordinates, and variant text type. The distribution characteristics are convolved using a preset convolution kernel trained on historical high-incidence scenes; finally, the scene attention weight is used as a channel attention mask and multiplied element-wise with the convolution operation output to generate a dynamic baseline value. Then, based on the dynamic baseline value and the screen features, the scene label vector in the current live broadcast screen features is extracted; the cosine similarity between the scene label vector and the historical scene label in the dynamic baseline value is calculated to generate the scene matching coefficient; finally, the maximum value in the dynamic baseline value is weighted summed with the scene matching coefficient to output the final bullet-message trigger weight.

[0093] This solution determines the frequency of occurrence of abnormal barrages and variant text in the corresponding distributed screen based on the barrage distribution characteristics, helping to quantify the spatiotemporal distribution density of different areas of the screen. Based on the frequency of occurrence, a scene attention weight is generated to help characterize the risk level of malicious content in the current scene. Convolution operations are performed on the distribution characteristics and the scene attention weight to output a dynamic baseline value, which helps to respond to feedback lag optimization needs. Based on the dynamic baseline value, the live screen characteristics are analyzed and the barrage trigger weights of the live screen characteristics are determined, which helps to achieve the core goal of adaptive risk management.

[0094] In some embodiments, an initial tolerance threshold is determined based on the live broadcast type; the rationality of the occurrence of abnormal barrage is determined based on the barrage trigger weight; historical reporting data is obtained, analyzed, and a negative feedback threshold is determined; the initial tolerance threshold is adjusted based on the rationality of occurrence and the negative feedback threshold to obtain a dynamic shielding judgment threshold for abnormal barrage.

[0095] The initial tolerance threshold may be a reference value for the frequency of abnormal barrages that is preset based on the live broadcast type.

[0096] The rationality of occurrence can be the result of determining whether the abnormal barrage is consistent with the current live broadcast scenario based on the barrage trigger weight.

[0097] Historical reporting data can be user reporting records of the same live broadcast type within a specified time period.

[0098] The negative feedback threshold may be a dynamic judgment standard generated based on historical reporting data.

[0099] Specifically, the live broadcast type is identified based on the live broadcast screen features, and the initial tolerance threshold of the corresponding type is retrieved from the preset rule library. Then, the barrage trigger weight is input into the piecewise linear function to generate the rationality coefficient of appearance, thereby determining the rationality of the appearance of abnormal barrage. Subsequently, the number of barrage reports per unit time within the same live broadcast type in history is counted, and the percentile value is calculated as the negative feedback baseline threshold; then, the reporting data at the current moment is weighted to generate a negative feedback threshold. The real-time risk situation represented by the rationality of appearance and the group feedback sensitivity represented by the negative feedback threshold are jointly regulated, and a dynamic weighted correction is performed on the initial tolerance threshold to generate a dynamic shielding judgment threshold for abnormal barrage that adapts to the risk level of the current live broadcast scenario.

[0100] This solution determines the initial tolerance threshold based on the live broadcast type, helping to eliminate the problem of poor scene adaptability and avoid misjudgments caused by a one-size-fits-all approach. Determining the legitimacy of abnormal barrages based on their trigger weights helps quickly suppress malicious screen-swiping behavior and avoid excessive interference with benign interactions. Historical reporting data is obtained and analyzed to determine the negative feedback threshold and block the risk of negative content spreading. Adjusting the initial tolerance threshold based on the legitimacy of occurrence and the negative feedback threshold yields a dynamic blocking threshold for abnormal barrages, helping to maintain basic functionality in extreme situations and avoid the risk of system crashes caused by delayed feedback.

[0101] In some embodiments, the live screen features are analyzed to determine the visual feature vector and the audio keyword vector; the visual feature vector and the audio keyword vector are cross-modally fused to generate a scene semantic vector; abnormal barrages and variant texts are analyzed to determine the text semantic vector; the dynamic cosine similarity between the text semantic vector and the scene semantic vector is calculated; the dynamic cosine similarity is matched with the dynamic baseline value, and based on the matching result, the barrage trigger weight of the live screen features is determined.

[0102] The visual feature vector can be a numerical representation extracted from the live broadcast picture.

[0103] The audio keyword vector may be a word embedding vector of entity keywords extracted from a live audio stream.

[0104] The scene semantic vector may be a unified semantic vector that represents the contextual semantics of the current live broadcast scene.

[0105] The text semantic vector can be a contextual semantic representation.

[0106] The dynamic cosine similarity may be a cosine similarity used to quantify the relevance between the bullet comment text and the live broadcast scene.

[0107] The matching result may be a comparison relationship between the dynamic cosine similarity and the dynamic baseline value.

[0108] Specifically, a visual feature extraction model pre-trained on large-scale image datasets such as ImageNet is used to perform frame sampling and parsing of live broadcast images through transfer learning, generating visual feature vectors representing the main objects and actions in the image. Simultaneously, an audio keyword recognition engine based on an LSTM-HMM architecture extracts entity keywords from the live audio stream and generates audio keyword vectors after word embedding mapping. Then, a dual-channel attention gating mechanism is constructed, modeled on the Transformer multi-head attention mechanism. A cross-modal gating unit is introduced to calculate the correlation weight matrix between visual and audio features. Cross-modal attention fusion is performed on the audio keyword vector based on the correlation weight matrix to output a scene semantic vector. A TextCNN model optimized through adversarial training is used to map abnormal comments and variant text into text semantic vectors. Dynamic baseline values ​​are dynamically generated based on a statistical model that correlates historical user negative feedback rates, comment density, and image features. Based on the content category identifier of the current live stream, the corresponding typed dynamic baseline values ​​are loaded from a preset rule library. The scene semantic vectors are mapped to the text semantic space using a trainable projection matrix to generate an alignment vector. The dynamic cosine similarity between the text semantic vector and the alignment vector is calculated. The dynamic cosine similarity is matched with the dynamic baseline value in multiple intervals. According to the matching results, a piecewise linear function is used to determine the bullet screen trigger weight of the live screen feature.

[0109] This solution analyzes live video features to determine visual feature vectors and audio keyword vectors, helping to eliminate the lack of correlation between video features and comment content. It also captures the potential correlation between audio semantics and comment content, avoiding misjudgments caused by independently processing modal data. Cross-modal attention fusion of visual feature vectors and audio keyword vectors generates scene semantic vectors, helping to establish a joint semantic space for video, audio, and comment content, providing a baseline scene context for dynamic similarity calculation. It analyzes anomalous comment content and variant text to determine text semantic vectors, capturing implicit malicious intent in comment content and avoiding mechanical misjudgments caused by keyword matching. It calculates the dynamic cosine similarity between text semantic vectors and scene semantic vectors, dynamically identifying scene mismatch risks and eliminating poor scene adaptability. The dynamic cosine similarity is matched with a dynamic baseline value. Based on the matching results, the comment trigger weight for live video features is determined, enabling dynamic control of threshold rigidity and feedback lag, rapidly improving blocking sensitivity.

[0110] In some embodiments, a semantic derivation path of malicious semantic deformation features is simulated by a generative adversarial network; and a malicious semantic feature matrix is ​​constructed based on the semantic derivation path.

[0111] A generative adversarial network can be a deep learning architecture consisting of a generator and a discriminator.

[0112] The semantic derivation path can be the evolution trajectory of malicious text from original semantics to deformed semantics in the latent space of the adversarial generative network.

[0113] Specifically, a generative adversarial network (GAN) was used to input normal bullet comment text into the generator. Combined with a cascaded three-layer LSTM network architecture, the lexical replacement layer, syntactic reorganization layer, and pragmatic adaptation layer of the text were modeled, respectively. Hierarchical gradient transfer was used to generate malicious semantic deformation features. Furthermore, a high-dimensional semantic manifold was constructed within the generator's latent space. Semantic derivative paths were then extracted on this high-dimensional semantic manifold through Monte Carlo sampling. Based on the semantic derivative paths, the SimHash algorithm was used to hash the path nodes, calculate the Hamming distance change rate, and generate a locally sensitive hash value. The JS divergence between the path endpoint and starting point was calculated to determine the semantic drift. The locally sensitive hash value and semantic drift were then integrated to construct a malicious semantic feature matrix.

[0114] This solution uses a generative adversarial network to simulate the semantic derivation path of malicious semantic deformation features, providing a mathematical representation of the dynamic deformation trajectory for feature matrix construction and supporting the abstract extraction of semantic deformation patterns. Constructing a malicious semantic feature matrix based on the semantic derivation path helps meet the technical requirements for semantic deformation feature mining.

[0115] Figure 3 This is a structural diagram of a digital intelligent live broadcast bullet screen monitoring system provided in one embodiment of the present application, such as Figure 3 As shown, the digital intelligent live broadcast barrage supervision system 300 of this embodiment includes: a data analysis module 301, a threshold determination module 302, and a strategy determination module 303.

[0116] The data analysis module 301 is used to obtain live broadcast barrage data and real-time live broadcast data; analyze the live broadcast barrage data, determine abnormal barrages and the abnormality level of the abnormal barrages; A threshold determination module 302 is configured to perform a risk assessment on the abnormal barrage based on the real-time live broadcast data and determine a dynamic blocking threshold for the abnormal barrage; The strategy determination module 303 is used to compare the abnormality level with the dynamic blocking determination threshold, and determine the display strategy of the abnormal barrage according to the level comparison result.

[0117] Optionally, when the data analysis module 301 analyzes the live broadcast barrage data and determines abnormal barrages and the abnormality level of the abnormal barrages, it is used to: Analyze the real-time live broadcast data to determine the live broadcast type; Based on the live broadcast type, obtain a preset sensitive word library; Based on the preset sensitive word library, analyzing the live broadcast barrage data to determine abnormal barrage; Analyze the abnormal comments and determine the degree of deviation of values; The abnormality level of the abnormal barrage is determined according to the value deviation.

[0118] Optionally, the data analysis module 301 analyzes the live broadcast barrage data based on the preset sensitive word library, and when determining abnormal barrage, is used to: Based on the real-time live broadcast data, historical live broadcast barrage data is obtained; Determining malicious semantic deformation features based on the preset sensitive word library and the historical live broadcast barrage data; Constructing a malicious semantic feature matrix based on the malicious semantic deformation features; Mapping the live broadcast barrage data to the malicious semantic feature matrix through the BERT model to obtain a multi-dimensional matrix value; Abnormal comments are determined according to the multi-dimensional matrix value.

[0119] Optionally, when the data analysis module 301 analyzes the abnormal comments and determines the value deviation, it is used to: Obtaining historical cyber violence events, analyzing the historical cyber violence events, and determining text feature distribution patterns; According to the live broadcast type, obtain the supervision level setting parameters of the current live broadcast room; Determine a preset multi-dimensional warning threshold based on the text feature distribution law and the supervision level setting parameters; The multi-dimensional matrix value of the abnormal barrage is compared with the preset multi-dimensional warning threshold, and the value deviation is determined according to the warning comparison result.

[0120] Optionally, the threshold determination module 302 performs a risk assessment on the abnormal barrage in combination with the real-time live broadcast data, and determines a dynamic blocking determination threshold for the abnormal barrage, for: Analyze the abnormal comments according to the malicious semantic feature matrix to determine variant text; Analyze the real-time live broadcast data to determine the live broadcast image characteristics at the current moment; Acquire historical real-time live broadcast data of the current live broadcast room from the start of the live broadcast to the current moment, analyze the historical real-time live broadcast data, and determine the distribution characteristics of the abnormal barrage and the variant text barrage; Determining the bullet screen trigger weight of the live broadcast screen feature according to the bullet screen distribution feature; The barrage trigger weight is used to perform a risk assessment on the abnormal barrage, and a dynamic shielding judgment threshold for the abnormal barrage is determined.

[0121] Optionally, when the threshold determination module 302 determines the bullet chat trigger weight of the live screen feature according to the bullet chat distribution feature, it is configured to: Determining the frequency of occurrence of the abnormal barrage and the variant text in the corresponding distribution screen according to the barrage distribution characteristics; Generate scene attention weights based on the occurrence frequencies; Performing a convolution operation on the distribution feature and the scene attention weight to output a dynamic baseline value; Based on the dynamic baseline value, the live broadcast screen features are analyzed to determine the bullet screen triggering weight of the live broadcast screen features.

[0122] Optionally, the threshold determination module 302 uses the barrage triggering weight to perform a risk assessment on the abnormal barrage, and when determining the dynamic shielding determination threshold for the abnormal barrage, is used to: Determining an initial tolerance threshold according to the live broadcast type; Determining the rationality of the occurrence of the abnormal barrage based on the barrage trigger weight; Obtaining historical reporting data, analyzing the historical reporting data, and determining a negative feedback threshold; According to the rationality of the occurrence and the negative feedback threshold, the initial tolerance threshold is adjusted to obtain the dynamic shielding determination threshold of the abnormal barrage.

[0123] Optionally, when the threshold determination module 302 analyzes the live screen features based on the dynamic baseline value and determines the bullet screen triggering weight of the live screen features, it is configured to: Analyze the live broadcast image features to determine visual feature vectors and audio keyword vectors; Performing cross-modal attention fusion on the visual feature vector and the audio keyword vector to generate a scene semantic vector; Parsing the abnormal comment and the variant text to determine a text semantic vector; Calculating the dynamic cosine similarity between the text semantic vector and the scene semantic vector; The dynamic cosine similarity is matched with the dynamic baseline value, and the bullet screen triggering weight of the live screen feature is determined according to the matching result.

[0124] Optionally, when constructing a malicious semantic feature matrix based on the malicious semantic deformation feature, the data analysis module 301 is configured to: Simulating the semantic derivation path of the malicious semantic deformation feature through a generative adversarial network; According to the semantic derivation path, a malicious semantic feature matrix is ​​constructed.

[0125] The system of this embodiment can be used to execute the method of any of the above embodiments. Its implementation principles and technical effects are similar and will not be described in detail here.

Claims

1. A digital intelligent live broadcast bullet screen supervision method, characterized in that: include: Get live broadcast barrage data and real-time live broadcast data; Analyze the live broadcast barrage data to determine abnormal barrages and the abnormality level of the abnormal barrages; In combination with the real-time live broadcast data, a risk assessment is performed on the abnormal barrage, and a dynamic blocking determination threshold for the abnormal barrage is determined; The abnormality level is compared with the dynamic shielding determination threshold, and the display strategy of the abnormal barrage is determined according to the level comparison result.

2. The method according to claim 1, characterized in that The analyzing the live broadcast barrage data to determine abnormal barrages and the abnormality level of the abnormal barrages includes: Analyze the real-time live broadcast data to determine the live broadcast type; Based on the live broadcast type, obtain a preset sensitive word library; Based on the preset sensitive word library, analyzing the live broadcast barrage data to determine abnormal barrage; Analyze the abnormal comments and determine the degree of deviation of values; The abnormality level of the abnormal barrage is determined according to the value deviation.

3. The method according to claim 2, characterized in that The analyzing the live broadcast barrage data based on the preset sensitive word library to determine abnormal barrage includes: Based on the real-time live broadcast data, historical live broadcast barrage data is obtained; Determining malicious semantic deformation features based on the preset sensitive word library and the historical live broadcast barrage data; Constructing a malicious semantic feature matrix based on the malicious semantic deformation features; Mapping the live broadcast barrage data to the malicious semantic feature matrix through the BERT model to obtain a multi-dimensional matrix value; Abnormal comments are determined according to the multi-dimensional matrix value.

4. The method according to claim 3, characterized in that The analyzing the abnormal comments and determining the value deviation includes: Obtaining historical cyber violence events, analyzing the historical cyber violence events, and determining text feature distribution patterns; According to the live broadcast type, obtain the supervision level setting parameters of the current live broadcast room; Determine a preset multi-dimensional warning threshold based on the text feature distribution law and the supervision level setting parameters; The multi-dimensional matrix value of the abnormal barrage is compared with the preset multi-dimensional warning threshold, and the value deviation is determined according to the warning comparison result.

5. The method according to claim 3, characterized in that The step of performing a risk assessment on the abnormal barrage in combination with the real-time live broadcast data and determining a dynamic blocking threshold for the abnormal barrage includes: Analyze the abnormal comments according to the malicious semantic feature matrix to determine variant text; Analyze the real-time live broadcast data to determine the live broadcast image characteristics at the current moment; Acquire historical real-time live broadcast data of the current live broadcast room from the start of the live broadcast to the current moment, analyze the historical real-time live broadcast data, and determine the distribution characteristics of the abnormal barrage and the variant text barrage; Determining the bullet screen trigger weight of the live broadcast screen feature according to the bullet screen distribution feature; The barrage trigger weight is used to perform a risk assessment on the abnormal barrage, and a dynamic shielding judgment threshold for the abnormal barrage is determined.

6. The method according to claim 5, characterized in that The determining, according to the bullet screen distribution characteristics, the bullet screen triggering weight of the live screen characteristics includes: Determining the frequency of occurrence of the abnormal barrage and the variant text in the corresponding distribution screen according to the barrage distribution characteristics; Generate scene attention weights based on the occurrence frequencies; Performing a convolution operation on the distribution feature and the scene attention weight to output a dynamic baseline value; Based on the dynamic baseline value, the live broadcast screen features are analyzed to determine the bullet screen triggering weight of the live broadcast screen features.

7. The method according to claim 5, characterized in that The step of utilizing the barrage trigger weight to perform risk assessment on the abnormal barrage and determining a dynamic shielding determination threshold for the abnormal barrage includes: Determining an initial tolerance threshold according to the live broadcast type; Determining the rationality of the occurrence of the abnormal barrage based on the barrage trigger weight; Obtaining historical reporting data, analyzing the historical reporting data, and determining a negative feedback threshold; According to the rationality of the occurrence and the negative feedback threshold, the initial tolerance threshold is adjusted to obtain the dynamic shielding determination threshold of the abnormal barrage.

8. The method according to claim 6, characterized in that The analyzing the live screen feature based on the dynamic baseline value and determining the bullet screen triggering weight of the live screen feature includes: Analyze the live broadcast image features to determine visual feature vectors and audio keyword vectors; Performing cross-modal attention fusion on the visual feature vector and the audio keyword vector to generate a scene semantic vector; Parsing the abnormal comment and the variant text to determine a text semantic vector; Calculating the dynamic cosine similarity between the text semantic vector and the scene semantic vector; The dynamic cosine similarity is matched with the dynamic baseline value, and the bullet screen triggering weight of the live screen feature is determined according to the matching result.

9. The method according to claim 3, characterized in that The step of constructing a malicious semantic feature matrix based on the malicious semantic deformation feature includes: Simulating the semantic derivation path of the malicious semantic deformation feature through a generative adversarial network; According to the semantic derivation path, a malicious semantic feature matrix is ​​constructed.

10. A digital intelligent live broadcast bullet screen monitoring system, characterized in that: The method as claimed in any one of claims 1 to 9 comprises: A data analysis module is used to obtain live broadcast barrage data and real-time live broadcast data; analyze the live broadcast barrage data, determine abnormal barrages and the abnormality level of the abnormal barrages; A threshold determination module is used to perform a risk assessment on the abnormal barrage in combination with the real-time live broadcast data, and determine a dynamic shielding determination threshold for the abnormal barrage; The strategy determination module is used to compare the abnormal level with the dynamic shielding judgment threshold, and determine the display strategy of the abnormal barrage according to the level comparison result.

Images